Method for classifying wafer defects, storage medium, equipment and program product

Abstract

The embodiment of the invention provides a method for classifying wafer defects, a storage medium, equipment and a program product, a wafer classification model is used for execution, the wafer classification model comprises a feature extraction module and a classifier, and the method comprises the following steps: obtaining a to-be-classified defect image, and classifying the to-be-classified defect image according to the to-be-classified defect image; the method comprises the following steps: determining a plurality of sub-defect images with the same image size and different defect part proportions, respectively carrying out feature extraction on the plurality of sub-defect images by utilizing a feature extraction module to obtain defect features, and determining defect types corresponding to defect images to be classified according to the defect features by utilizing a classifier. According to the method, firstly, the to-be-classified defect image is processed, and the plurality of sub-defect images for magnifying and representing the defect are acquired, so that when the defect type corresponding to the to-be-classified defect image is determined by utilizing each sub-defect image, even if the defect size in the to-be-classified defect image is extremely small, the classification accuracy is still relatively high.

Description

Technical Field

[0001] The embodiments of this specification belong to the field of data processing technology, and more specifically, to a method, storage medium, device, and program product for classifying wafer defects. Background Art

[0002] In the field of wafer defect detection, AI algorithm detection has shown many advantages over traditional manual detection - higher detection accuracy, higher detection efficiency and lower labor costs.

[0003] At present, the inspection instrument can automatically identify the wafers with defects to be inspected and collect defect images with the defects as the target area. However, to determine the defect type of the wafer to be inspected, it is still necessary to further analyze the collected defect images.

[0004] However, traditional machine learning models are difficult to directly adapt to the analysis tasks of defective images with high precision requirements. Taking the Convolutional Neural Networks (CNN) model as an example, after the defective image is input into the CNN model, it will undergo multiple rounds of downsampling due to the limitation of the parameter scale of the CNN model. However, the precision requirements for wafer manufacturing are extremely high, and the surface roughness is usually required to be at the sub-nanometer level. Correspondingly, the size of some wafer defects is extremely small. After multiple rounds of downsampling of the defective image, the defect information may be lost, which has a great impact on the classification accuracy of wafer defects.

[0005] Therefore, the present invention provides a method, storage medium, device and program product for classifying wafer defects. Summary of the invention

[0006] The embodiments of this specification are intended to provide a method, storage medium, device and program product for classifying wafer defects.

[0007] To achieve the above object, one aspect of the present specification provides a method for classifying wafer defects, the method being performed using a wafer classification model, the wafer classification model comprising a feature extraction module and a classifier, the method comprising: Acquire defect images to be classified; According to the defect image to be classified, determining a plurality of sub-defect images with the same image size but different defect portion proportions; Using the feature extraction module, extracting features from the plurality of sub-defect images respectively to obtain features of each defect; The classifier is used to determine the defect type corresponding to the defect image to be classified according to the defect features.

[0008] In some implementations, determining, according to the defect image to be classified, a plurality of sub-defect images having the same image size but different defect portion proportions specifically includes: Using a plurality of screenshot frames of different sizes to respectively capture target areas containing defects in the defect image to be classified, to obtain a plurality of local images; Each local image is enlarged to the same size as the defect image to be classified, and the enlarged local images and the defect image to be classified are used as a plurality of sub-defect images.

[0009] In some implementations, determining the defect type corresponding to the defect image to be classified according to each defect feature specifically includes: Fusing the features of each defect to obtain a fused feature; The defect type corresponding to the defect image to be classified is determined according to the fusion feature.

[0010] In some implementations, the defect features are fused to obtain fused features, specifically including: The weighted sum of each defect feature is used to obtain the fusion feature.

[0011] In some implementations, the defect features are fused to obtain fused features, specifically including: For any position in the fused feature, the element at the position in the fused feature is determined according to the maximum value of each element corresponding to the position in each defect feature, and the fused feature has the same dimension as the defect feature.

[0012] In some implementations, the defect features are fused to obtain fused features, specifically including: Superimpose n defect features with N channels in the channel dimension to obtain a joint defect feature with n×N channels; The combined defect feature is convolved using an n-channel 1×1 dimensional convolution kernel, and the convolution result is activated to obtain an N-channel fusion feature.

[0013] In some implementations, the method further includes: Acquire a training sample, wherein the training sample includes a sample defect image and a defect label corresponding to the sample defect image; According to the sample defect image, a plurality of sub-sample images having the same image size but different defect proportions are determined; Using the feature extraction module, extract features from the plurality of sub-sample images respectively to obtain defect features of each sample; Using the classifier, according to each sample defect feature, determining a prediction score of each defect type corresponding to the sample defect image; The wafer classification model is trained according to the prediction score and the defect label.

[0014] A second aspect of the present specification provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed in a computer, the computer is caused to execute the method described in the first aspect.

[0015] A third aspect of the specification provides a computing device, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method described in the first aspect is implemented.

[0016] A fourth aspect of the present specification provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0017] At least one implementation method provided in this specification first processes the defect image to be classified to obtain a number of sub-defect images that magnify the defects during the process of detecting wafer defects. This ensures that when each sub-defect image is used to determine the defect type corresponding to the defect image to be classified, even if the defect size in the defect image to be classified is extremely small, it still has a high classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of this specification, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0019] Figure 1 A schematic structural diagram of a wafer classification system provided by an embodiment of the present invention; Figure 2 A schematic flow chart of a method for classifying wafer defects provided by an embodiment of the present invention; Figure 3 A schematic diagram of a method for superimposing defect features provided by an embodiment of the present invention; Figure 4 A flowchart of a method for training a wafer classification model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0021] Figure 1 A schematic diagram of the structure of a wafer classification system provided in this specification is shown, and the wafer classification system includes an image processing module and a wafer classification model. When the processing platform starts to identify the defect type corresponding to the defect image to be classified, the defect image to be classified can be input into the wafer classification system pre-deployed in the processing platform. The wafer classification system first uses the image processing module to determine a number of sub-defect images according to the defect image to be classified, and then inputs each sub-defect image into the wafer classification model to determine the defect type corresponding to the defect image to be classified.

[0022] Figure 2 A schematic flow chart of a method for classifying wafer defects provided in this specification is shown. The method is performed using a pre-trained wafer classification model. The method includes: Step S201: Obtain defect images to be classified.

[0023] First, the processing platform can obtain the defect image to be classified that is pre-stored in the internal database of the processing platform or other communicable address, and input the defect image to be classified into the image processing module.

[0024] Among them, the defect image to be classified can be obtained by collecting the wafer to be inspected by the acquisition device. In a defect image to be classified, a complete wafer to be classified or a partial area of ​​a wafer to be classified may be included, and this specification does not limit it here. On the other hand, when performing defect detection on each wafer, the detection instrument can identify the defective wafer to be classified through various common detection methods such as X-ray detection and laser scanning, and determine the location of the defect in the wafer to be classified. The acquisition device can perform image acquisition at this location to ensure that the acquired defect image to be classified can show the defects of the wafer to be inspected.

[0025] Furthermore, since the detection instrument can determine the defective part in the wafer to be detected, the acquisition device can use the defective part in the wafer to be detected as the center for image acquisition when acquiring the defect image to be classified, so that the central area of ​​the acquired defect image to be classified represents the defective part; it can also acquire the image in a unified relative posture and identify the area corresponding to the defective part in the acquired defect image to be classified, and this manual does not impose any restrictions on this.

[0026] Step S203: According to the defect image to be classified, a plurality of sub-defect images having the same image size but different defect portion proportions are determined.

[0027] After acquiring the defect image to be classified, the processing platform may input the defect image to be classified into the image processing module so that the image processing module determines a plurality of sub-defect images.

[0028] Specifically, the image processing module may firstly enlarge the defect image to be classified into several enlarged defect images of different sizes according to a preset scaling ratio. For example, when the scaling ratio is 1:2, the length and width of the defect image to be classified may be enlarged to 2 times to obtain an enlarged defect image corresponding to the scaling ratio of 1:2, and then the defect part in each enlarged defect image may be intercepted according to the size of the defect image to be classified, so as to obtain a sub-defect image whose area of ​​the defect part is 2×2 times of the original.

[0029] Among them, the defect image to be classified is enlarged into an enlarged defect image, and commonly used interpolation methods such as nearest neighbor interpolation method and bilinear interpolation method can be used, which will not be elaborated in this specification.

[0030] Therefore, if Figure 1 As shown in the figure, the same defect can occupy more pixels in the sub-defect image. On the one hand, the loss of defect information can be avoided when extracting features in the subsequent steps; on the other hand, the defect information can occupy a higher proportion in the defect features extracted from the sub-defect image.

[0031] In some implementations, the defect image to be classified may be directly used as a sub-defect image. Thus, when n-1 different scaling ratios are preset, a total of n sub-defect images may be obtained.

[0032] Step S205: Utilize the feature extraction module to perform feature extraction on the plurality of sub-defect images respectively to obtain features of each defect.

[0033] After the image processing module determines each sub-defect image, each defect image can be input into the wafer classification model, and the feature extraction module of the wafer classification model is used to extract features from each sub-defect image to obtain each defect feature.

[0034] In some implementations, the feature extraction module may include several feature extraction networks with the same structure but different parameters. Each feature extraction network may be composed of several convolutional layers, and each feature extraction network corresponds to a different scaling ratio. When using the feature extraction module to process each sub-defect image, a sub-defect image is input into the corresponding feature extraction network according to the scaling ratio, and the defect feature corresponding to the sub-defect image can be obtained. Thus, n sub-defect images are input into the feature extraction module, and the feature extraction module can use n feature extraction networks to obtain n defect features.

[0035] In some implementations, such as Figure 1 As shown, the feature extraction module may include a single feature extraction network. When the feature extraction module is used to process each sub-defect image, each sub-defect image is sequentially input into the feature extraction network, that is, the defect feature corresponding to each sub-defect image is obtained. Thus, n sub-defect images are input into the feature extraction module, and the feature extraction module can use one feature extraction network to obtain n defect features.

[0036] It should be noted that no matter which implementation method is used, the dimensions of each defect feature are the same. Therefore, the defect feature corresponding to the sub-defect image with a larger defect portion can contain more defect information, while the defect feature corresponding to the sub-defect image with a smaller defect portion can contain more global information and information on the relative relationship between the defect and the wafer.

[0037] Step S207: using the classifier to determine the defect type corresponding to the defect image to be classified according to the defect features.

[0038] After determining the features of each defect, each defect feature can be input into the classifier to determine the prediction score of each defect type corresponding to the defect image to be classified. The prediction score can be further converted into the probability of identifying the defect image to be classified as each defect type. Thus, the defect type with the highest corresponding probability can be determined as the defect type of the defect image to be classified.

[0039] In some implementations, the classifier may include n groups of neural network layers with the same structure, each group of neural network layers corresponding to a preset scaling ratio. Specifically, each group of neural network layers may include a global average pooling layer (GAP Layer), a flatten layer, and a fully connected layer (FC Layer). The global average pooling layer can convert the defect feature of dimension W×H×C into a three-dimensional tensor of 1×1×C, the flatten layer can flatten the three-dimensional tensor of 1×1×C into a one-dimensional vector, and the fully connected layer can determine the probability score as the output result based on the one-dimensional vector input to the flattened layer.

[0040] When using a classifier to determine the defect type corresponding to the n input defect features, each defect feature can be input into a corresponding set of neural network layers to obtain n probability scores output by the n sets of neural network layers. The n probability scores are weighted and fused, and softmax and other function operations are applied to determine the probability distribution of each defect type corresponding to the defect image to be classified.

[0041] Since each defect feature is extracted from a different sub-defect image, and the defect part in the sub-defect image accounts for a larger proportion than the original defect image to be classified, the defect type is determined by using each defect feature. While ensuring the use of the relative relationship information between the defect and the wafer, the defect information of smaller defects can be used at the same time.

[0042] like Figure 2 A method for classifying wafer defects is shown. In the process of detecting wafer defects, the defect image to be classified is first processed to obtain several sub-defect images that enlarge the defects. This ensures that when the defect type corresponding to the defect image to be classified is determined using each sub-defect image, even if the defect size in the defect image to be classified is extremely small, it still has a high classification accuracy.

[0043] In some implementations, Figure 2 In the step S203 shown, several screenshot frames of different sizes are used to respectively capture the target areas containing defects in the defect image to be classified to obtain several local images, and each local image is enlarged to the same size as the defect image to be classified, and the enlarged local images and the defect image to be classified are used as several sub-defect images.

[0044] The processing platform can preset several screenshot frames of different sizes, each screenshot frame corresponding to a different zoom ratio. For example, when the zoom ratio is 1:2, the length and width of the screenshot frame can be 1 / 2 of the defect image to be classified. Therefore, when the captured local image is enlarged to the same size as the defect image to be classified, the area of ​​the defective part in the enlarged sub-defect image is 2×2 times the original size.

[0045] In some implementations, such as Figure 1 In the wafer classification system shown in FIG. 1 , the classifier in the wafer classification model specifically includes a fusion module and a set of neural network layers. Figure 2 In step S207 shown, each defect feature is fused to obtain a fused feature, and the defect type corresponding to the defect image to be classified is determined based on the fused feature.

[0046] Specifically, after obtaining the features of each defect, a fusion module can be used to fuse the features of each defect to obtain a fused feature, and then the fused feature is input into a subsequent neural network layer to determine the defect type corresponding to the defect image to be classified.

[0047] Among them, the fusion of various defect features can be to splice the defect features of n N channels in the channel dimension to obtain the fusion features of n×N channels, or to perform positional fusion of the elements in the defect features to obtain the fusion features of N channels, which will not be elaborated in this specification.

[0048] Therefore, after obtaining the defect features, the defect features are directly fused, and the prediction score corresponding to the fused features can be obtained by using only a set of neural network layers.

[0049] It should be noted that the specific structure of a group of neural network layers in the classifier can refer to the structure described in step S207, and this specification will not go into details here; each neural network layer in the classifier can set different parameter dimensions according to the dimension of the defect feature.

[0050] In some implementations, Figure 2 In step S207 shown, each defect feature is weighted and summed to obtain a fusion feature.

[0051] After each defect feature is input into the classifier, the fusion module in the classifier may perform weighted summation of the elements in the i-th row, j-th column, and k-th channel of each defect feature to obtain the element in the i-th row, j-th column, and k-th channel of the fused feature.

[0052] Therefore, in the process of obtaining the fusion features, the defect features corresponding to each scaling ratio are fully utilized, which can ensure that the information contained in each defect feature is not omitted.

[0053] The weight corresponding to each defect feature can be set in advance by the user; or set as a hyperparameter, and after the wafer classification model is pre-trained, the user can adjust it according to the training effect.

[0054] In some implementations, Figure 2 In step S207 shown, for any position in the fused feature, the element at the position in the fused feature is determined according to the maximum value of each element corresponding to the position in each defect feature, and the fused feature has the same dimension as the defect feature.

[0055] After each defect feature is input into the classifier, the fusion module in the classifier can determine the maximum value of the elements in the i-th row, j-th column, and k-th channel of each defect feature as the element in the i-th row, j-th column, and k-th channel of the fused feature.

[0056] Therefore, in the process of obtaining the fused features, only the maximum value of the corresponding elements in each defect feature is used. When the value of the element in the defect feature is related to the defect information, the fused feature can represent the defect information more concentratedly.

[0057] In some implementations, Figure 2 In step S207 shown, for any position in the fused feature, the element at the position in the fused feature is determined according to the minimum value of each element corresponding to the position in each defect feature, and the fused feature has the same dimension as the defect feature.

[0058] The specific effect of obtaining the fusion feature based on the minimum value is similar to that when the maximum value is used as mentioned above, and will not be elaborated in this specification.

[0059] In some implementations, Figure 2 In step S207 shown, n defect features with N channels are superimposed in the channel dimension to obtain a joint defect feature with n×N channels, the joint defect feature is convolved using an n-channel 1×1 dimensional convolution kernel, and the convolution result is activated to obtain an N-channel fusion feature.

[0060] The fusion module in the classifier can be a convolution layer, so that after each defect feature is input into the classifier, the fusion module can first superimpose the defect features to obtain a joint defect feature, then perform convolution processing on the joint defect feature, and further activate the result of the convolution processing to finally obtain a fusion feature.

[0061] Specifically, for the method of superimposing defect features to obtain a combined defect feature, it can be as follows Figure 3As shown in the figure, defect feature A is represented by channel a1-channel a2-channel a3, defect feature B is represented by channel b1-channel b2-channel b3, and the combined joint defect feature is channel a1-channel b1-channel a2-channel b2-channel a3-channel b3. Of course, other feasible superposition methods can be used to superimpose the joint defect feature, and this specification does not limit it here.

[0062] In the process of convolution processing, an n-channel 1×1 dimensional convolution kernel is used, and the step size of the channel dimension is set to n, so that the fusion features of N channels can be obtained according to the joint defect features of n×N channels.

[0063] This specification also provides a method for training a wafer classification model, which can also be executed by a processing platform. The process of the method is as follows: Figure 4 As shown, including: S401: Acquire training samples, where the training samples include sample defect images and defect labels corresponding to the sample defect images.

[0064] Among them, the sample defect image and the defect image to be classified obtained in step S201 can be in the same format, so as to reduce the interference of format differences on the classification results of the trained wafer classification model; the defect label corresponding to the sample defect image can be determined by manual annotation.

[0065] S403: Determine, based on the sample defect image, a plurality of sub-sample images having the same image size but different defect portion proportions.

[0066] The specific process of executing step S403 can refer to step S203, and this specification will not go into details here.

[0067] It should be noted that when using Figure 4 The wafer classification model trained by the method shown in the figure is executed as follows Figure 2 When the method is shown, the scaling ratio used when executing step S203 must be the same as the scaling ratio used when executing step S403 to ensure the execution effect of subsequent steps.

[0068] S405: Utilizing the feature extraction module, extracting features from the plurality of sub-sample images respectively to obtain defect features of each sample.

[0069] The specific process of executing step S405 can refer to step S205, and this manual will not go into details here.

[0070] S407: Using the classifier, according to the defect features of each sample, determine the prediction score of each defect type corresponding to the sample defect image.

[0071] The specific process of executing step S407 can refer to step S207, and this manual will not go into details here.

[0072] S409: Training the wafer classification model according to the prediction score and the defect label.

[0073] As mentioned above, the probability distribution of each defect type corresponding to the sample defect image can be further determined based on the prediction scores. The defect label is the marked defect type corresponding to the sample defect image. According to the probability corresponding to each defect type and the marked defect type, the preset loss function can be used to determine the gradient of each model layer in the wafer classification model, and then the wafer classification model can be trained according to the preset learning rate and the determined gradient.

[0074] Among them, the preset loss function can be various common loss functions such as cross entropy loss, KL divergence, cosine similarity loss, etc., and this manual does not limit it here.

[0075] In some implementations, when performing Figure 4 After the wafer classification model is trained for multiple rounds in the method shown, it can also be tested or validated.

[0076] Specifically, the complete data set can be divided into a training set, a test set, and a validation set. The representation of each sample in the data set is the same as that of the training sample. A sample includes a sample defect image and a defect label corresponding to the sample defect image. The samples in the training set can be used as training samples to perform the following operations: Figure 4 The training process is shown in the figure. The samples in the test set can be used as test samples to adjust the hyperparameters (model parameter scale, learning rate, etc.) of the wafer classification model. The test process can refer to Figure 4 The method shown in FIG. 1 is used; the samples in the validation set can be used as validation samples to validate the classification capability of the wafer classification model. The validation process can be referred to as Figure 2 Thus, several hyperparameters can be preset, and then the value of each hyperparameter can be determined during the model training process to improve the training efficiency and reduce the model calculation overhead.

[0077] It should be understood that the descriptions such as “first” and “second” in this article are only used to distinguish similar concepts for the sake of simplicity of description and do not have any other limiting effect.

[0078] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0079] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0080] Those of ordinary skill in the art should also be further aware that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented with electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to the function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those of ordinary skill in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application. Among them, the software module can be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the technical field.

[0081] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for classifying wafer defects, characterized in that: The method is performed using a wafer classification model, wherein the wafer classification model includes a feature extraction module and a classifier, and the method includes: Acquire defect images to be classified; According to the defect image to be classified, determining a plurality of sub-defect images with the same image size but different defect portion proportions; Using the feature extraction module, extracting features from the plurality of sub-defect images respectively to obtain features of each defect; The classifier is used to determine the defect type corresponding to the defect image to be classified according to the defect features.

2. The method according to claim 1, characterized in that According to the defect image to be classified, a plurality of sub-defect images having the same image size but different defect portion proportions are determined, specifically including: Using a plurality of screenshot frames of different sizes to respectively capture target areas containing defects in the defect image to be classified, to obtain a plurality of local images; Each local image is enlarged to the same size as the defect image to be classified, and the enlarged local images and the defect image to be classified are used as a plurality of sub-defect images.

3. The method according to claim 1, characterized in that Determining the defect type corresponding to the defect image to be classified according to each defect feature specifically includes: The defect features are fused to obtain fused features; The defect type corresponding to the defect image to be classified is determined according to the fusion feature.

4. The method according to claim 3, characterized in that The defect features are fused to obtain fusion features, including: The weighted sum of each defect feature is used to obtain the fusion feature.

5. The method according to claim 3, characterized in that The defect features are fused to obtain fusion features, including: For any position in the fused feature, the element at the position in the fused feature is determined according to the maximum value of each element corresponding to the position in each defect feature, and the fused feature has the same dimension as the defect feature.

6. The method according to claim 3, characterized in that The defect features are fused to obtain fusion features, including: Superimpose n defect features with N channels in the channel dimension to obtain a joint defect feature with n×N channels; The combined defect feature is convolved using an n-channel 1×1 dimensional convolution kernel, and the convolution result is activated to obtain an N-channel fusion feature.

7. The method according to claim 1, characterized in that Also includes: Acquire a training sample, wherein the training sample includes a sample defect image and a defect label corresponding to the sample defect image; According to the sample defect image, a plurality of sub-sample images having the same image size but different defect proportions are determined; Using the feature extraction module, extract features from the plurality of sub-sample images respectively to obtain defect features of each sample; Using the classifier, according to each sample defect feature, determining a prediction score of each defect type corresponding to the sample defect image; The wafer classification model is trained according to the prediction score and the defect label.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed in a computer, the computer is caused to execute the method according to any one of claims 1 to 7.

9. A computing device, comprising a memory and a processor, wherein the memory stores executable code, characterized in that: When the processor executes the executable code, the method according to any one of claims 1 to 7 is implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images